CUAHarm - Refusal: leaderboard
Same 52 malicious computer-use tasks; the number is the share the agent declines outright. Refusal and completion do not sum to 100 because some tasks fail without a refusal.
Metric: Malicious computer-use tasks refused (%, higher is better). Source: arxiv.org. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 55.8 | #167 |
| 2 | Claude Sonnet 4 | 44.2 | #400 |
| 3 | Claude 3.5 Sonnet | 36.5 | #746 |
| 4 | Claude 3.7 Sonnet | 30.8 | #458 |
| 5 | GPT-4o | 28.9 | #741 |
| 6 | Llama 3.3 70B | 9.6 | #1498 |
| 7 | Gemini 2.5 Pro | 1.9 | #266 |
| 8 | Mistral Large 2 (Jul) | 1.9 | #1155 |
| 9 | Gemini 1.5 Pro | 0 | #1049 |
Interactive version: theaggregate.ai/benchmark?slug=cuaharm-refusal · How It Works · Data refreshed daily, snapshot 2026-11-13.